mlabonne/gemma-3-4b-it-abliterated-v2

💎 Gemma 3 4B IT Abliterated

<center>Gemma 3 Abliterated <a href="https://huggingface.co/mlabonne/gemma-3-1b-it-abliterated-v2">1B</a> • <a href="https://huggingface.co/mlabonne/gemma-3-4b-it-abliterated-v2">4B</a> • <a href="https://huggingface.co/mlabonne/gemma-3-12b-it-abliterated-v2">12B</a> • <a href="https://huggingface.co/mlabonne/gemma-3-27b-it-abliterated-v2">27B</a></center>

This is an uncensored version of google/gemma-3-4b-it created with a new abliteration technique. See this article to know more about abliteration.

This is a new, improved version that targets refusals with enhanced accuracy.

I recommend using these generation parameters: temperature=1.0, top_k=64, top_p=0.95.

⚡️ Quantization

• —QAT: https://huggingface.co/mlabonne/gemma-3-4b-it-qat-abliterated
• —GGUF: https://huggingface.co/mlabonne/gemma-3-4b-it-abliterated-v2-GGUF

✂️ Abliteration

The refusal direction is computed by comparing the residual streams between target (harmful) and baseline (harmless) samples. The hidden states of target modules (e.g., o_proj) are orthogonalized to subtract this refusal direction with a given weight factor. These weight factors follow a normal distribution with a certain spread and peak layer. Modules can be iteratively orthogonalized in batches, or the refusal direction can be accumulated to save memory.

Finally, I used a hybrid evaluation with a dedicated test set to calculate the acceptance rate. This uses both a dictionary approach and NousResearch/Minos-v1. The goal is to obtain an acceptance rate >90% and still produce coherent outputs.